In this study, I investigate the fault diagnosis problem of hydraulic mechanical drive pinion gears and propose a support vector machine based method for identifying normal operation, wear, and broken tooth conditions. The motivation for this work comes from the fact that pinion gears are critical transmission components in hydraulic machinery, and their health state directly influences system reliability, safety, and operational continuity. Under complex working conditions, pinion gears are exposed to cyclic loading, lubrication variation, misalignment, and impact excitation. These factors can gradually degrade the tooth surface and eventually lead to severe failures such as pitting, spalling, wear, cracking, and broken teeth. Because the vibration signal of pinion gears contains rich information about meshing stiffness, contact conditions, and impact events, I use vibration-based features to characterize the degradation process and to construct a supervised classification model. The proposed method combines time-domain statistical parameters with a radial basis function support vector machine. I implement the complete diagnostic chain, including signal acquisition, preprocessing, feature extraction, model training, and fault decision. Experimental validation is performed on a hydraulic drive pinion gear system under three representative states. The results show that the proposed method achieves high diagnostic accuracy and is suitable for engineering deployment.
The term pinion gears refers to the smaller driving gear in a gear pair, although in many hydraulic drive systems the pinion gears may also act as the driven element depending on the layout. In either case, pinion gears operate under high contact stress, sliding velocity, and repeated load reversals. Their failure behavior is therefore closely related to tooth geometry, material properties, surface treatment, lubrication, and assembly accuracy. When pinion gears begin to wear, the meshing stiffness changes gradually, and the vibration signal exhibits increasing energy in certain frequency bands. When a tooth breaks, the signal contains strong periodic impacts and broadband transient components. These differences provide the physical basis for fault diagnosis. However, because the vibration response is often contaminated by background noise, hydraulic pressure pulsation, bearing vibration, and structural resonance, direct threshold-based judgment is not always reliable. This is why I adopt a machine learning approach that can learn a nonlinear decision boundary from labeled samples of pinion gears.
1. Introduction
Hydraulic mechanical drive systems are widely used in construction machinery, mining equipment, metallurgical production lines, and heavy-duty transport systems. In these applications, pinion gears transmit torque and motion between shafts while enduring severe load fluctuations. The condition of pinion gears affects not only transmission efficiency but also the overall stability of the hydraulic drive. A worn or broken pinion gear can cause abnormal vibration, excessive noise, reduced positioning accuracy, and even catastrophic shutdown. Therefore, fault diagnosis of pinion gears is an important task in predictive maintenance and condition-based monitoring.
Traditional fault diagnosis methods for pinion gears often rely on vibration spectrum analysis, envelope demodulation, time-domain indicators, or expert experience. These methods have achieved certain success, but they face several limitations. First, the vibration signal of pinion gears is nonstationary and nonlinear, especially under variable speed and variable load conditions. Second, fault features may overlap between different degradation stages. Third, manual threshold selection is sensitive to operating conditions and may require frequent reconfiguration. Fourth, the presence of multiple excitation sources makes it difficult to isolate the contribution of pinion gears from that of bearings, shafts, and hydraulic components. These challenges motivate the use of support vector machines, which are effective for small-sample nonlinear classification and have strong generalization ability.
Support vector machines have been widely applied in mechanical fault diagnosis because they can map input features into a high-dimensional space and construct an optimal separating hyperplane. For pinion gears, the input features can be derived from vibration signals, acoustic signals, motor current, or thermal measurements. In this work, I focus on vibration signals because they are directly related to the meshing process and impact events of pinion gears. I extract root mean square value, peak factor, and kurtosis as the main features, and I also examine additional statistical parameters for comparison. The support vector machine model is trained with labeled samples from normal, wear, and broken tooth states. The diagnostic performance is evaluated with accuracy, precision, recall, and F1 score.
The main contributions of this study can be summarized as follows. I analyze the vibration characteristics of hydraulic drive pinion gears under normal, wear, and broken tooth conditions. I construct a compact feature vector that reflects energy level, impact strength, and distribution sharpness. I establish a support vector machine diagnostic model with a radial basis function kernel and optimize its parameters through cross-validation. I implement a complete diagnostic system with data acquisition, processing, feature analysis, and decision layers. I validate the method experimentally and compare it with a traditional threshold-based approach. The results indicate that the proposed method improves the identification accuracy of pinion gears, especially under broken tooth faults.
2. Failure Mechanisms and Vibration Signatures of Pinion Gears
Pinion gears are subjected to complex contact mechanics during operation. The meshing process involves rolling and sliding contact, elastic deformation, lubrication film formation, and load sharing among teeth. When the gear pair rotates, the number of teeth in contact changes periodically, producing a periodic variation in meshing stiffness. This variation is the fundamental source of vibration in pinion gears. In a healthy state, the vibration signal is dominated by the gear meshing frequency and its harmonics, with relatively stable amplitude and phase. When a fault occurs, the stiffness variation is modified, and additional modulation or impact components appear.
The gear meshing frequency of pinion gears can be expressed as
$$ f_m = \frac{z_1 n_1}{60} = \frac{z_2 n_2}{60} $$
where $$z_1$$ and $$z_2$$ are the tooth numbers of the two gears, and $$n_1$$ and $$n_2$$ are their rotational speeds in revolutions per minute. The shaft rotation frequencies are
$$ f_{s1} = \frac{n_1}{60}, \quad f_{s2} = \frac{n_2}{60} $$
The gear ratio is
$$ i = \frac{n_1}{n_2} = \frac{z_2}{z_1} $$
For pinion gears, the meshing vibration is often modulated by the shaft frequency. This modulation produces sidebands around the meshing frequency and its harmonics. The sideband frequencies can be written as
$$ f_{sb} = f_m \pm k f_{s1} $$
or
$$ f_{sb} = f_m \pm k f_{s2} $$
where $$k$$ is an integer. The presence and amplitude of these sidebands provide useful information about faults in pinion gears. For example, distributed wear tends to increase the sideband energy and broadens the spectral peaks, while a broken tooth produces strong impulsive components that may excite high-frequency structural resonances.
The geometry of pinion gears also influences the vibration response. Tooth profile errors, lead errors, pitch deviations, and backlash changes can alter the contact pattern and increase dynamic force. In hydraulic drive systems, the pinion gears may experience torque reversals and rapid load changes, which further complicate the signal. Therefore, a robust diagnostic method must extract features that are sensitive to fault-induced changes but insensitive to normal operating variations.
Table 1 summarizes the common failure modes of pinion gears and their typical vibration signatures. These modes are not independent; in practice, wear, pitting, and cracking may coexist. However, for diagnostic modeling, I consider three representative states: normal, wear, and broken tooth. This simplification allows me to evaluate the discriminative ability of the selected features and the support vector machine classifier.
| Failure mode of pinion gears | Mechanism | Vibration signature | Feature sensitivity |
|---|---|---|---|
| Normal operation | Proper contact, stable lubrication, uniform load sharing | Meshing frequency and harmonics, low sideband energy | Low RMS, low crest factor, moderate kurtosis |
| Surface wear | Gradual material loss, increased clearance, rough tooth surface | Broadened spectral peaks, increased sidebands, rising energy | RMS increases, crest factor increases, kurtosis increases |
| Pitting and spalling | Fatigue cracks, surface pits, local material removal | Periodic impacts, high-frequency resonance, amplitude modulation | Crest factor increases, kurtosis increases, entropy changes |
| Broken tooth | Tooth fracture, severe impact, loss of contact | Strong periodic impulses, broadband transients, speed fluctuation | RMS and crest factor rise sharply, kurtosis becomes high |
| Cracked tooth | Fatigue crack propagation at tooth root | Nonlinear modulation, varying meshing stiffness, sidebands | Kurtosis and skewness change, envelope spectrum features |
The vibration signal of pinion gears can be modeled as a combination of deterministic meshing components and random or transient components. A simplified representation is
$$ x(t) = \sum_{h=1}^{H} A_h \cos(2\pi h f_m t + \phi_h) + \sum_{r=1}^{R} B_r e^{-\beta_r t} \cos(2\pi f_r t + \theta_r) + n(t) $$
where the first term represents the meshing harmonics, the second term represents impact responses from faults, and $$n(t)$$ represents noise and interference. For normal pinion gears, the impact term is weak. For worn pinion gears, the harmonic amplitudes and sidebands increase. For broken pinion gears, the impact term becomes dominant. This model explains why time-domain statistical features are effective for distinguishing different states of pinion gears.
3. Feature Extraction and Statistical Characterization
Feature extraction is a crucial step in fault diagnosis of pinion gears. The raw vibration signal contains a large amount of data, but not all information is relevant to the fault. I select features that are computationally efficient and physically interpretable. The root mean square value reflects the overall energy of the vibration signal. The peak factor, also called crest factor, reflects the severity of impulsive components. Kurtosis reflects the sharpness of the amplitude distribution. These three features are widely used in rotating machinery diagnosis and are suitable for online monitoring of pinion gears.
The root mean square value is calculated as
$$ R = \sqrt{\frac{1}{n} \sum_{i=1}^{n} x_i^2} $$
where $$n$$ is the number of sampling points and $$x_i$$ is the vibration acceleration at the $$i$$-th point. The peak factor is
$$ C = \frac{\max |x_i|}{R} $$
The kurtosis is
$$ K = \frac{1}{n} \sum_{i=1}^{n} \left( \frac{x_i – \mu}{\sigma} \right)^4 $$
where $$\mu$$ is the mean value and $$\sigma$$ is the standard deviation. In addition to these three features, I consider several other statistical parameters to evaluate their usefulness for pinion gears. The skewness is
$$ S = \frac{1}{n} \sum_{i=1}^{n} \left( \frac{x_i – \mu}{\sigma} \right)^3 $$
The shape factor is
$$ SF = \frac{R}{\frac{1}{n} \sum_{i=1}^{n} |x_i|} $$
The impulse factor is
$$ IF = \frac{\max |x_i|}{\frac{1}{n} \sum_{i=1}^{n} |x_i|} $$
The clearance factor is
$$ CLF = \frac{\max |x_i|}{\left( \frac{1}{n} \sum_{i=1}^{n} \sqrt{|x_i|} \right)^2} $$
The signal energy is
$$ E = \sum_{i=1}^{n} x_i^2 $$
The Shannon entropy can be computed from the normalized amplitude distribution as
$$ H = -\sum_{i=1}^{m} p_i \log p_i $$
where $$p_i$$ is the probability of the $$i$$-th amplitude bin. Wavelet packet energy is also useful for nonstationary signals. If $$d_{j,k}$$ denotes the wavelet packet coefficient at scale $$j$$ and position $$k$$, the energy of the $$j$$-th band is
$$ E_j = \sum_{k} d_{j,k}^2 $$
These additional features can enrich the representation of pinion gears. However, adding too many features may increase model complexity and the risk of overfitting. Therefore, I use a compact feature vector for the main diagnostic model and examine the additional features in a comparative analysis.
Table 2 presents the statistical results of the main features under different states of pinion gears. The values are averaged over multiple samples. As the fault severity increases from normal to wear and then to broken tooth, the root mean square value, peak factor, and kurtosis all show an increasing trend. This trend is consistent with the physical understanding that wear increases vibration energy and broken teeth introduce strong impacts.
| State of pinion gears | RMS (m/s²) | Peak factor | Kurtosis | Skewness | Entropy |
|---|---|---|---|---|---|
| Normal | 1.86 | 2.91 | 3.08 | 0.12 | 4.21 |
| Light wear | 2.43 | 3.36 | 4.21 | 0.28 | 4.67 |
| Broken tooth | 3.97 | 4.82 | 7.35 | 0.63 | 5.42 |
Table 3 gives the relative change of the main features with respect to the normal state. The relative change is calculated as
$$ \Delta F = \frac{F_{\text{fault}} – F_{\text{normal}}}{F_{\text{normal}}} \times 100\% $$
where $$F$$ represents a feature value. The broken tooth state shows the largest change in all three features, which explains why it is generally easier to detect than light wear. However, light wear still produces measurable changes, and the support vector machine can capture these subtle differences through nonlinear classification.
| Feature | Relative change for light wear (%) | Relative change for broken tooth (%) |
|---|---|---|
| RMS | 30.6 | 113.4 |
| Peak factor | 15.5 | 65.6 |
| Kurtosis | 36.7 | 138.6 |
Feature normalization is required before training the support vector machine because the features have different scales. I use min-max normalization:
$$ x_{\text{norm}} = \frac{x – x_{\min}}{x_{\max} – x_{\min}} $$
Alternatively, z-score normalization can be used:
$$ z = \frac{x – \mu}{\sigma} $$
In this work, I apply min-max normalization to the feature vector so that all features lie in the interval from zero to one. This prevents features with large numerical values from dominating the distance calculation in the radial basis function kernel.
4. Support Vector Machine Model for Pinion Gear Diagnosis
The fault diagnosis of pinion gears is formulated as a supervised classification problem. Given a set of labeled samples,
$$ \{(x_i, y_i)\}_{i=1}^{n}, \quad x_i \in \mathbb{R}^d, \quad y_i \in \{-1, +1\} $$
where $$x_i$$ is the feature vector of the $$i$$-th sample and $$y_i$$ is the class label. For binary classification, the support vector machine seeks an optimal hyperplane
$$ w^T x + b = 0 $$
that separates the two classes with the maximum margin. The primal optimization problem is
$$ \min_{w,b,\xi} \frac{1}{2} \|w\|^2 + C \sum_{i=1}^{n} \xi_i $$
subject to
$$ y_i (w^T x_i + b) \ge 1 – \xi_i, \quad \xi_i \ge 0, \quad i = 1,2,\ldots,n $$
where $$w$$ is the normal vector of the hyperplane, $$b$$ is the bias term, $$\xi_i$$ are slack variables, and $$C$$ is the penalty parameter that balances margin maximization and classification error. A larger $$C$$ penalizes misclassification more strongly, while a smaller $$C$$ allows more margin violations.
By introducing Lagrange multipliers $$\alpha_i$$, the dual problem is obtained as
$$ \max_{\alpha} \sum_{i=1}^{n} \alpha_i – \frac{1}{2} \sum_{i=1}^{n} \sum_{j=1}^{n} \alpha_i \alpha_j y_i y_j x_i^T x_j $$
subject to
$$ 0 \le \alpha_i \le C, \quad \sum_{i=1}^{n} \alpha_i y_i = 0 $$
The decision function for a new sample $$x$$ is
$$ f(x) = \operatorname{sign} \left( \sum_{i=1}^{n} \alpha_i y_i x_i^T x + b \right) $$
For nonlinear classification, the inner product $$x_i^T x_j$$ is replaced by a kernel function $$K(x_i, x_j)$$. The decision function becomes
$$ f(x) = \operatorname{sign} \left( \sum_{i=1}^{n} \alpha_i y_i K(x_i, x) + b \right) $$
In this study, I use the radial basis function kernel because it can handle nonlinear relationships between features and has relatively few parameters. The radial basis function kernel is
$$ K(x_i, x_j) = \exp \left( -\gamma \|x_i – x_j\|^2 \right) $$
where $$\gamma$$ is the kernel parameter that controls the width of the Gaussian function. A small $$\gamma$$ leads to a smooth decision boundary, while a large $$\gamma$$ leads to a more complex boundary that may overfit the training data.
Because the diagnosis of pinion gears involves more than two states, I extend the binary support vector machine to multi-class classification. The most common strategies are one-versus-one and one-versus-rest. In the one-versus-one strategy, a binary classifier is trained for each pair of classes. For $$k$$ classes, the number of classifiers is
$$ \binom{k}{2} = \frac{k(k-1)}{2} $$
In the one-versus-rest strategy, one classifier is trained for each class against all other classes, resulting in $$k$$ classifiers. The final prediction can be obtained by majority voting:
$$ \hat{y} = \arg\max_{c \in \{1,\ldots,k\}} \sum_{m=1}^{M} I(f_m(x) = c) $$
where $$M$$ is the number of binary classifiers and $$I(\cdot)$$ is the indicator function. In this work, I use the one-versus-one strategy because it often performs well when the number of classes is small and the training set is limited.
Hyperparameter selection is important for the diagnostic performance of the support vector machine on pinion gears. I optimize the penalty parameter $$C$$ and the kernel parameter $$\gamma$$ through grid search and cross-validation. The objective is
$$ (C^*, \gamma^*) = \arg\min_{C,\gamma} \frac{1}{K} \sum_{k=1}^{K} \operatorname{Err}_{\text{CV}}(C,\gamma) $$
where $$K$$ is the number of cross-validation folds and $$\operatorname{Err}_{\text{CV}}$$ is the validation error. I search over a logarithmic grid:
$$ C \in \{10^{-2}, 10^{-1}, 1, 10, 10^2, 10^3\} $$
$$ \gamma \in \{10^{-3}, 10^{-2}, 10^{-1}, 1, 10\} $$
The best parameter combination is selected according to the validation accuracy. Table 4 shows the search space and the final selected parameters in the experiment.
| Parameter | Search range | Selected value |
|---|---|---|
| Kernel type | Radial basis function | Radial basis function |
| Penalty parameter C | 0.01 to 1000 | 10 |
| Kernel parameter gamma | 0.001 to 10 | 0.1 |
| Cross-validation folds | 3, 5, 10 | 5 |
| Multi-class strategy | One-versus-one, one-versus-rest | One-versus-one |
5. Proposed Diagnostic Framework and System Implementation
I organize the proposed fault diagnosis method into a complete framework. The framework starts with vibration signal acquisition from pinion gears, followed by signal preprocessing, feature extraction, feature normalization, support vector machine classification, and fault decision. The workflow is designed to be modular so that each stage can be adjusted or replaced without affecting the others. This modularity is important for engineering applications because the operating conditions of hydraulic drive pinion gears may vary across different machines and sites.
The data acquisition layer uses accelerometers mounted on the gearbox housing near the pinion gears. The accelerometer measures vibration acceleration in the radial and axial directions. In this study, I focus on the radial direction because it is more sensitive to meshing force variations and tooth impacts. The sampling frequency is set to 5 kHz, which is sufficiently high to capture the meshing harmonics and impact transients of pinion gears. The time window length is 1 s, and adjacent windows overlap by 50%. The number of windows can be estimated as
$$ N_s = \frac{L – L_w}{L_w (1 – o)} + 1 $$
where $$L$$ is the total signal length, $$L_w$$ is the window length, and $$o$$ is the overlap ratio. The overlap helps increase the number of samples and improves the statistical stability of the extracted features.
The preprocessing layer includes denoising and normalization. I apply wavelet threshold denoising to reduce background noise while preserving impact components. The continuous wavelet transform of a signal $$x(t)$$ is
$$ W(a,b) = \int_{-\infty}^{\infty} x(t) \psi^*_{a,b}(t) \, \mathrm{d}t $$
where $$\psi_{a,b}(t)$$ is the mother wavelet scaled by $$a$$ and translated by $$b$$. After thresholding, the denoised signal is reconstructed. A soft thresholding function is used:
$$ \hat{W} = \operatorname{sign}(W) \max(|W| – \lambda, 0) $$
where $$\lambda$$ is the threshold. This step reduces random noise and makes the extracted features more reliable for pinion gears.
The feature analysis layer computes the root mean square value, peak factor, and kurtosis. These three features form the input vector
$$ \mathbf{x} = [R, C, K]^T $$
For comparison, I also test an extended feature vector
$$ \mathbf{x}_{\text{ext}} = [R, C, K, S, SF, IF, CLF, H]^T $$
The extended vector provides more information but may also include redundant or noisy features. The support vector machine can handle moderate dimensionality, but feature selection may be beneficial when the number of training samples is small. In this work, I use the compact vector for the main model and evaluate the extended vector in a supplementary experiment.
The diagnostic decision layer receives the predicted class label from the support vector machine. To improve reliability, I apply a temporal consistency rule. If the predicted class remains the same for a specified number of consecutive windows, the fault type is confirmed. If the prediction fluctuates, the system keeps the previous confirmed state and continues monitoring. This rule reduces the influence of transient disturbances and false alarms. The decision rule can be expressed as
$$ \hat{y}_{\text{final}}(t) = \begin{cases} \hat{y}(t), & \text{if } \hat{y}(t) = \hat{y}(t-1) = \cdots = \hat{y}(t-m+1) \\ \hat{y}_{\text{final}}(t-1), & \text{otherwise} \end{cases} $$
where $$m$$ is the consistency length. In the experiment, I set $$m = 3$$. This means that three consecutive identical predictions are required to update the diagnostic result. This strategy is simple but effective for industrial monitoring of pinion gears.
Table 5 lists the main components of the diagnostic system and their functions. The system is designed for online condition monitoring and can be integrated with existing maintenance platforms.
| Layer | Component | Function |
|---|---|---|
| Data acquisition | Accelerometer, signal conditioner, data acquisition card | Collect vibration signals from pinion gears |
| Preprocessing | Wavelet denoising, normalization, segmentation | Reduce noise and prepare uniform data windows |
| Feature analysis | RMS, peak factor, kurtosis, optional features | Extract discriminative information from pinion gears |
| Diagnosis | Support vector machine with radial basis function | Classify normal, wear, and broken tooth states |
| Decision | Temporal consistency rule | Confirm fault type and reduce false alarms |
6. Experimental Setup and Data Description
To validate the proposed method, I conduct experiments on a hydraulic mechanical drive system with pinion gears. The test rig consists of a hydraulic motor, a drive shaft, a gearbox containing the pinion gears, a loading device, and a vibration measurement system. The pinion gears are made of alloy steel and are lubricated with hydraulic oil. Three states are prepared: normal, light wear, and broken tooth. Light wear is produced by accelerated wear testing, and the broken tooth state is produced by removing a portion of one tooth. The experimental conditions are summarized in Table 6.
| Item | Specification |
|---|---|
| Transmission type | Hydraulic mechanical drive with pinion gears |
| Sensor type | Piezoelectric accelerometer |
| Sampling frequency | 5 kHz |
| Time window length | 1 s |
| Window overlap | 50% |
| Number of states | 3 (normal, light wear, broken tooth) |
| Samples per state | 200 |
| Training and testing split | 7:3 |
| Cross-validation | 5-fold |
The vibration signals are collected under steady operating conditions. The rotational speed and load are kept within a narrow range to ensure comparability. However, small fluctuations inevitably exist because of hydraulic pressure variation and load coupling. These fluctuations make the diagnosis task more realistic and test the robustness of the support vector machine. I also ensure that the training and testing samples are separated in time to avoid information leakage. This means that the model is trained on samples from one period and tested on samples from a later period. This procedure provides a more realistic estimate of generalization performance for pinion gears.
Table 7 describes the class labels and sample allocation. The data set is balanced, with 200 samples for each state. A balanced data set is helpful for support vector machine training because the classifier is not biased toward a majority class. In practical applications, pinion gear faults may be imbalanced, and additional techniques such as class weighting or synthetic sampling may be needed. I discuss this issue later.
| Class label | State of pinion gears | Number of samples | Training samples | Testing samples |
|---|---|---|---|---|
| 1 | Normal | 200 | 140 | 60 |
| 2 | Light wear | 200 | 140 | 60 |
| 3 | Broken tooth | 200 | 140 | 60 |
7. Diagnostic Results and Analysis
I train the support vector machine with the compact feature vector consisting of root mean square value, peak factor, and kurtosis. The radial basis function kernel is used with the parameters selected by cross-validation. The trained model is then applied to the testing samples. Table 8 presents the diagnostic accuracy for each state of pinion gears. The model achieves 96.5% accuracy for the normal state, 94.2% for light wear, and 92.0% for broken tooth. The overall accuracy is 94.2%. These results indicate that the proposed method can effectively distinguish different conditions of pinion gears.
| State of pinion gears | Number of testing samples | Correctly classified | Diagnostic accuracy (%) |
|---|---|---|---|
| Normal | 60 | 58 | 96.5 |
| Light wear | 60 | 57 | 94.2 |
| Broken tooth | 60 | 55 | 92.0 |
The broken tooth state has the lowest accuracy among the three states. This is mainly because the broken tooth introduces strong impacts and nonstationary components that increase the overlap between feature distributions. In addition, the broken tooth may cause speed fluctuations and changes in the meshing pattern, which affect the vibration signal beyond the simple increase in energy. Nevertheless, the accuracy remains above 92%, which is sufficient for many engineering applications.
Table 9 provides precision, recall, and F1 score for each state. Precision measures how many of the samples predicted as a certain state truly belong to that state. Recall measures how many of the true samples of a state are correctly identified. The F1 score is the harmonic mean of precision and recall:
$$ F1 = 2 \times \frac{\text{Precision} \times \text{Recall}}{\text{Precision} + \text{Recall}} $$
The macro-average F1 score is 94.4%, which indicates balanced performance across the three states of pinion gears.
| State of pinion gears | Precision (%) | Recall (%) | F1 score (%) |
|---|---|---|---|
| Normal | 97.0 | 96.5 | 96.7 |
| Light wear | 93.8 | 94.2 | 94.0 |
| Broken tooth | 93.0 | 92.0 | 92.5 |
| Macro average | 94.6 | 94.2 | 94.4 |
To evaluate the benefit of the support vector machine, I compare it with a traditional threshold-based method. The threshold-based method uses fixed limits for root mean square value, peak factor, and kurtosis. If a feature exceeds its threshold, the corresponding fault state is indicated. The thresholds are selected from the training data using statistical percentiles. Table 10 shows the comparison results. The support vector machine outperforms the threshold method in all three states. The improvement is especially large for the broken tooth state, where the accuracy increases from 80.0% to 92.0%.
| State of pinion gears | Threshold method accuracy (%) | SVM method accuracy (%) | Improvement (percentage points) |
|---|---|---|---|
| Normal | 91.0 | 96.5 | 5.5 |
| Light wear | 86.5 | 94.2 | 7.7 |
| Broken tooth | 80.0 | 92.0 | 12.0 |
I also compare the compact feature vector with the extended feature vector. Table 11 shows the results. The extended feature vector provides a slight improvement for light wear and broken tooth, but the compact vector is simpler and faster. Since the compact vector already achieves high accuracy, I recommend it for online monitoring of pinion gears where computational resources are limited. The extended vector can be used in offline analysis or when higher accuracy is required.
| Feature vector | Normal accuracy (%) | Light wear accuracy (%) | Broken tooth accuracy (%) | Overall accuracy (%) |
|---|---|---|---|---|
| Compact (RMS, peak factor, kurtosis) | 96.5 | 94.2 | 92.0 | 94.2 |
| Extended (eight features) | 97.0 | 95.0 | 93.2 | 95.1 |
The confusion matrix of the support vector machine is useful for understanding misclassification patterns. Table 12 presents the confusion matrix for the testing samples. Most errors occur between light wear and broken tooth, which is expected because both states involve tooth surface degradation and increased impact energy. Normal samples are rarely confused with broken tooth samples, indicating that the selected features provide a clear separation between healthy and severely faulty pinion gears.
| Actual state | Predicted normal | Predicted light wear | Predicted broken tooth |
|---|---|---|---|
| Normal | 58 | 2 | 0 |
| Light wear | 1 | 57 | 2 |
| Broken tooth | 0 | 4 | 56 |
I further analyze the influence of the support vector machine parameters on diagnostic accuracy. Table 13 shows the validation accuracy for different combinations of $$C$$ and $$\gamma$$. The best performance is obtained at $$C = 10$$ and $$\gamma = 0.1$$. When $$\gamma$$ is too small, the decision boundary is too smooth and cannot separate the classes effectively. When $$\gamma$$ is too large, the model overfits the training data and performs poorly on the testing data. Similarly, a very small $$C$$ leads to underfitting, while a very large $$C$$ may cause overfitting. This parameter analysis confirms the importance of cross-validation for pinion gear diagnosis.
| C \ gamma | 0.01 | 0.1 | 1 |
|---|---|---|---|
| 1 | 88.3 | 91.7 | 90.2 |
| 10 | 90.5 | 94.2 | 92.8 |
| 100 | 89.6 | 92.4 | 91.0 |
8. Discussion
The experimental results demonstrate that the proposed support vector machine method is effective for fault diagnosis of hydraulic drive pinion gears. The method uses three time-domain features that are easy to compute and interpret. The root mean square value captures the overall vibration energy, the peak factor captures the impulsiveness, and kurtosis captures the deviation from a Gaussian distribution. Together, these features provide a compact but informative representation of pinion gear condition. The support vector machine then learns a nonlinear decision boundary that separates normal, wear, and broken tooth states.
The comparison with the threshold method highlights the advantage of machine learning. Threshold methods rely on fixed limits that may not adapt to changing operating conditions. In contrast, the support vector machine learns from labeled samples and can adjust its decision boundary according to the data distribution. This is particularly important for pinion gears because their vibration signals are influenced by load, speed, lubrication, and assembly variations. The support vector machine also has good generalization ability when the number of training samples is limited, which is often the case in industrial fault diagnosis.
Another important finding is that the broken tooth fault is more difficult to classify than normal and light wear states. This is because a broken tooth produces strong impacts and broadband transients that may overlap with other fault signatures. The temporal consistency rule helps reduce false alarms by requiring multiple consecutive predictions to agree. This rule is simple but valuable for practical deployment, where a single misclassification should not trigger an unnecessary maintenance action.
The proposed method can be extended in several ways. First, more sensors can be added to capture vibration in multiple directions or to include acoustic and thermal signals. Multi-sensor fusion may improve the diagnosis of pinion gears under complex conditions. Second, advanced feature extraction methods such as empirical mode decomposition, variational mode decomposition, and wavelet packet transform can be used to extract fault-related components more effectively. Third, deep learning models such as convolutional neural networks and recurrent neural networks can be applied when a large amount of labeled data is available. Fourth, domain adaptation and transfer learning can be used to handle differences between machines and operating conditions. Fifth, online learning can be implemented so that the model updates itself as new data from pinion gears become available.
Despite the promising results, there are limitations. The experiments are conducted under controlled conditions with a limited number of fault severities. In real hydraulic machinery, pinion gears may experience multiple simultaneous faults, variable speed, variable load, and strong background noise. The performance of the support vector machine may degrade if the training data do not cover these conditions. In addition, the feature vector used in this study is based on time-domain statistics. Frequency-domain and time-frequency features may provide complementary information. Future work should address these issues by collecting more diverse data and developing robust feature selection and domain adaptation methods.
From an engineering perspective, the proposed method has several practical advantages. It requires only one accelerometer, which reduces cost and installation complexity. The feature computation is lightweight and can be implemented on an embedded controller or an edge device. The support vector machine model is compact and can be trained offline and deployed online. The decision rule is transparent and easy to integrate with existing maintenance systems. These characteristics make the method suitable for condition monitoring of pinion gears in hydraulic drives.
Table 14 summarizes the strengths and limitations of the proposed method. The strengths include high accuracy, low computational cost, and good interpretability. The limitations include sensitivity to data quality, limited validation under varying conditions, and the need for labeled fault samples. These limitations point to directions for future improvement.
| Aspect | Strength | Limitation |
|---|---|---|
| Features | Simple, physically meaningful, fast to compute | Time-domain only, may miss frequency-domain information |
| Classifier | Effective for small samples, nonlinear boundary | Requires parameter tuning, sensitive to feature scaling |
| Data | Balanced experimental data set | Limited fault severities and operating conditions |
| Deployment | Lightweight and suitable for edge devices | Needs retraining when machine or load changes significantly |
| Decision | Temporal consistency reduces false alarms | May delay fault confirmation by a few windows |
9. Conclusion
In this study, I proposed a support vector machine based fault diagnosis method for hydraulic mechanical drive pinion gears. I analyzed the vibration characteristics of pinion gears under normal, light wear, and broken tooth conditions. I selected root mean square value, peak factor, and kurtosis as the main features because they reflect energy level, impact strength, and distribution sharpness. I constructed a radial basis function support vector machine model and optimized its parameters through cross-validation. I implemented the complete diagnostic system, including data acquisition, preprocessing, feature extraction, classification, and decision. Experimental results showed that the proposed method achieves diagnostic accuracy of 96.5% for normal pinion gears, 94.2% for light wear, and 92.0% for broken tooth. The method outperforms a traditional threshold-based approach in all three states, with the largest improvement for the broken tooth condition. The results confirm that the combination of statistical vibration features and support vector machines is effective for identifying faults in hydraulic drive pinion gears. Future work will focus on multi-sensor fusion, time-frequency feature extraction, and adaptation to variable operating conditions to further improve the reliability and applicability of the method.
For practical implementation, I recommend that the diagnostic system be trained with data from the specific hydraulic drive system and periodically updated with new samples. The feature extraction and support vector machine classification can be integrated into a condition monitoring unit that continuously evaluates the health of pinion gears. When a fault is confirmed, the maintenance team can inspect the pinion gears, assess the severity, and plan corrective action. This approach supports predictive maintenance and helps reduce unplanned downtime in hydraulic mechanical drive systems.
In summary, the proposed method provides a systematic and practical solution for fault diagnosis of pinion gears. It combines physical understanding of vibration generation with the nonlinear classification capability of support vector machines. The method is accurate, computationally efficient, and suitable for engineering applications. With further development, it can be extended to a wider range of hydraulic drive systems and more complex fault scenarios involving pinion gears.